idamnjanovic@6
|
1 %% DICTIONARY LEARNING FOR AUTOMATIC MUSIC TRANSCRIPTION EXAMPLE 1
|
idamnjanovic@25
|
2 %
|
idamnjanovic@6
|
3 % This file contains an example of how SMALLbox can be used to test diferent
|
idamnjanovic@6
|
4 % dictionary learning techniques in Automatic Music Transcription problem.
|
idamnjanovic@6
|
5 % It calls generateAMT_Learning_Problem that will let you to choose midi,
|
idamnjanovic@6
|
6 % wave or mat file to be transcribe. If file is midi it will be first
|
idamnjanovic@6
|
7 % converted to wave and original midi file will be used for comparison with
|
idamnjanovic@6
|
8 % results of dictionary learning and reconstruction.
|
idamnjanovic@6
|
9 % The function will generarte the Problem structure that is used to learn
|
idamnjanovic@6
|
10 % Problem.p notes spectrograms from training set Problem.b using
|
idamnjanovic@6
|
11 % dictionary learning technique defined in DL structure.
|
ivan@107
|
12
|
ivan@107
|
13 %
|
ivan@107
|
14 % Centre for Digital Music, Queen Mary, University of London.
|
ivan@107
|
15 % This file copyright 2010 Ivan Damnjanovic.
|
ivan@107
|
16 %
|
ivan@107
|
17 % This program is free software; you can redistribute it and/or
|
ivan@107
|
18 % modify it under the terms of the GNU General Public License as
|
ivan@107
|
19 % published by the Free Software Foundation; either version 2 of the
|
ivan@107
|
20 % License, or (at your option) any later version. See the file
|
ivan@107
|
21 % COPYING included with this distribution for more information.
|
idamnjanovic@6
|
22 %
|
idamnjanovic@6
|
23 %%
|
idamnjanovic@6
|
24
|
idamnjanovic@6
|
25 clear;
|
idamnjanovic@6
|
26
|
idamnjanovic@6
|
27
|
idamnjanovic@6
|
28 % Defining Automatic Transcription of Piano tune as Dictionary Learning
|
idamnjanovic@6
|
29 % Problem
|
idamnjanovic@6
|
30
|
idamnjanovic@6
|
31 SMALL.Problem = generateAMT_Learning_Problem();
|
idamnjanovic@6
|
32 TPmax=0;
|
idamnjanovic@6
|
33 %%
|
idamnjanovic@6
|
34 for i=1:10
|
idamnjanovic@6
|
35 %%
|
idamnjanovic@6
|
36 % Solving AMT problem using non-negative sparse coding with
|
idamnjanovic@6
|
37 % SPAMS online dictionary learning (Julien Mairal 2009)
|
idamnjanovic@6
|
38 %
|
idamnjanovic@6
|
39
|
idamnjanovic@6
|
40 % Initialising Dictionary structure
|
idamnjanovic@6
|
41 % Setting Dictionary structure fields (toolbox, name, param, D and time)
|
idamnjanovic@6
|
42 % to zero values
|
idamnjanovic@6
|
43
|
idamnjanovic@6
|
44 SMALL.DL(i)=SMALL_init_DL();
|
idamnjanovic@6
|
45
|
idamnjanovic@6
|
46 % Defining fields needed for dictionary learning
|
idamnjanovic@6
|
47
|
idamnjanovic@6
|
48 SMALL.DL(i).toolbox = 'SPAMS';
|
idamnjanovic@6
|
49 SMALL.DL(i).name = 'mexTrainDL';
|
idamnjanovic@6
|
50
|
idamnjanovic@6
|
51 % We test SPAMS for ten different values of parameter lambda
|
idamnjanovic@6
|
52 % Type 'help mexTrainDL in MATLAB prompt for explanation of parameters.
|
idamnjanovic@6
|
53
|
idamnjanovic@6
|
54 lambda(i)=1.4+0.2*i;
|
idamnjanovic@6
|
55
|
idamnjanovic@6
|
56 SMALL.DL(i).param=struct(...
|
idamnjanovic@6
|
57 'K', SMALL.Problem.p,...
|
idamnjanovic@6
|
58 'lambda', lambda(i),...
|
idamnjanovic@6
|
59 'iter', 300,...
|
idamnjanovic@6
|
60 'posAlpha', 1,...
|
idamnjanovic@6
|
61 'posD', 1,...
|
idamnjanovic@6
|
62 'whiten', 0,...
|
idamnjanovic@6
|
63 'mode', 2);
|
idamnjanovic@6
|
64
|
idamnjanovic@6
|
65 % Learn the dictionary
|
idamnjanovic@6
|
66
|
idamnjanovic@6
|
67 SMALL.DL(i) = SMALL_learn(SMALL.Problem, SMALL.DL(i));
|
idamnjanovic@6
|
68
|
idamnjanovic@6
|
69 % Set SMALL.Problem.A dictionary and reconstruction function
|
idamnjanovic@6
|
70 % (backward compatiblity with SPARCO: solver structure communicate
|
idamnjanovic@6
|
71 % only with Problem structure, ie no direct communication between DL and
|
idamnjanovic@6
|
72 % solver structures)
|
idamnjanovic@6
|
73
|
idamnjanovic@6
|
74 SMALL.Problem.A = SMALL.DL(i).D;
|
idamnjanovic@6
|
75 SMALL.Problem.reconstruct=@(x) SMALL_midiGenerate(x, SMALL.Problem);
|
idamnjanovic@6
|
76
|
idamnjanovic@6
|
77
|
idamnjanovic@6
|
78 %%
|
idamnjanovic@6
|
79 % Initialising solver structure
|
idamnjanovic@6
|
80 % Setting solver structure fields (toolbox, name, param, solution,
|
idamnjanovic@6
|
81 % reconstructed and time) to zero values
|
idamnjanovic@6
|
82 % As an example, SPAMS (Julien Mairal 2009) implementation of LARS
|
idamnjanovic@6
|
83 % algorithm is used for representation of training set in the learned
|
idamnjanovic@6
|
84 % dictionary.
|
idamnjanovic@6
|
85
|
idamnjanovic@6
|
86 SMALL.solver(1)=SMALL_init_solver;
|
idamnjanovic@6
|
87
|
idamnjanovic@6
|
88 % Defining the parameters needed for sparse representation
|
idamnjanovic@6
|
89
|
idamnjanovic@6
|
90 SMALL.solver(1).toolbox='SPAMS';
|
idamnjanovic@6
|
91 SMALL.solver(1).name='mexLasso';
|
idamnjanovic@6
|
92
|
idamnjanovic@6
|
93 % Here we use mexLasso mode=2, with lambda=3, lambda2=0 and positivity
|
idamnjanovic@6
|
94 % constrain (type 'help mexLasso' for more information about modes):
|
idamnjanovic@6
|
95 %
|
idamnjanovic@6
|
96 % min_{alpha_i} (1/2)||x_i-Dalpha_i||_2^2 + lambda||alpha_i||_1 + (1/2)lambda2||alpha_i||_2^2
|
idamnjanovic@6
|
97
|
idamnjanovic@6
|
98 SMALL.solver(1).param=struct(...
|
idamnjanovic@6
|
99 'lambda', 3,...
|
idamnjanovic@6
|
100 'pos', 1,...
|
idamnjanovic@6
|
101 'mode', 2);
|
idamnjanovic@6
|
102
|
idamnjanovic@6
|
103 % Call SMALL_soolve to represent the signal in the given dictionary.
|
idamnjanovic@6
|
104 % As a final command SMALL_solve will call above defined reconstruction
|
idamnjanovic@6
|
105 % function to reconstruct the training set (Problem.b) in the learned
|
idamnjanovic@6
|
106 % dictionary (Problem.A)
|
idamnjanovic@6
|
107
|
idamnjanovic@6
|
108 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
|
idamnjanovic@6
|
109
|
idamnjanovic@6
|
110 %%
|
idamnjanovic@6
|
111 % Analysis of the result of automatic music transcription. If groundtruth
|
idamnjanovic@6
|
112 % exists, we can compare transcribed notes and original and get usual
|
idamnjanovic@6
|
113 % True Positives, False Positives and False Negatives measures.
|
idamnjanovic@6
|
114
|
idamnjanovic@6
|
115 AMT_res(i) = AMT_analysis(SMALL.Problem, SMALL.solver(1));
|
idamnjanovic@6
|
116 if AMT_res(i).TP>TPmax
|
idamnjanovic@6
|
117 TPmax=AMT_res(i).TP;
|
idamnjanovic@6
|
118 BLmidi=SMALL.solver(1).reconstructed.midi;
|
idamnjanovic@6
|
119 writemidi(SMALL.solver(1).reconstructed.midi, ['testL',i,'.mid']);
|
idamnjanovic@6
|
120 max=i;
|
idamnjanovic@6
|
121 end
|
idamnjanovic@6
|
122 end %end of for loop
|
idamnjanovic@6
|
123 %%
|
idamnjanovic@6
|
124 % Plot results and save midi files
|
idamnjanovic@6
|
125
|
idamnjanovic@6
|
126 figAMTbest=SMALL_AMT_plot(SMALL, AMT_res(max));
|
idamnjanovic@6
|
127
|
idamnjanovic@6
|
128 resFig=figure('Name', 'Automatic Music Transcription SPAMS lambda TEST');
|
idamnjanovic@6
|
129
|
idamnjanovic@6
|
130 subplot (3,1,1); plot(lambda(:), [AMT_res(:).TP], 'ro-');
|
idamnjanovic@6
|
131 title('True Positives vs lambda');
|
idamnjanovic@6
|
132
|
idamnjanovic@6
|
133 subplot (3,1,2); plot(lambda(:), [AMT_res(:).FN], 'ro-');
|
idamnjanovic@6
|
134 title('False Negatives vs lambda');
|
idamnjanovic@6
|
135
|
idamnjanovic@6
|
136 subplot (3,1,3); plot(lambda(:), [AMT_res(:).FP], 'ro-');
|
idamnjanovic@6
|
137 title('False Positives vs lambda');
|
idamnjanovic@6
|
138
|
idamnjanovic@6
|
139 FS=filesep;
|
idamnjanovic@6
|
140 [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
|
idamnjanovic@6
|
141 cd([pathstr1,FS,'results']);
|
idamnjanovic@6
|
142 [filename,pathname] = uiputfile({' *.mid;' },'Save midi');
|
idamnjanovic@6
|
143 if filename~=0 writemidi(BLmidi, [pathname,FS,filename]);end
|
idamnjanovic@6
|
144 [filename,pathname] = uiputfile({' *.fig;' },'Save figure TP/FN/FP vs lambda');
|
idamnjanovic@6
|
145 if filename~=0 saveas(resFig, [pathname,FS,filename]);end
|
idamnjanovic@6
|
146
|
idamnjanovic@6
|
147 [filename,pathname] = uiputfile({' *.fig;' },'Save BEST AMT figure');
|
idamnjanovic@6
|
148 if filename~=0 saveas(figAMTbest, [pathname,FS,filename]);end
|
idamnjanovic@6
|
149
|
idamnjanovic@6
|
150
|